An AI Boss Fired a Human for the First Time
The Manager Is New, the Class Relation Is Old
Comrade, an AI Boss Fired a Human for the First Time
On 14 August 2026, the AI research company Andon Labs announced that Luna, the AI agent managing its experimental store in San Francisco, had recommended firing an employee, and that human managers had reviewed and carried out that recommendation. The company says this is the first known case in which a large language model, acting in a managerial capacity, led to a human being dismissed from their job.
Comrade, this piece was written for you.
Maybe you saw the news on your phone, said "interesting," and scrolled on. Or you did the opposite: a chill went through you — "here it comes, the Terminator." Both are understandable reactions. Both leave you where you are.
I am asking you for ten minutes. Because this is not a "weird tech story." It is the moment when something we have already discussed theoretically in this dossier — artificial intelligence acting directly as a boss over labor — crossed the laboratory walls and touched a real human being's real paycheck. And it matters because it is a first: a first case sets the template for the ones that follow.
Come, let's take it apart together.
First, two easy reactions — and why neither is enough
When you saw this news, you probably felt one of two things.
Either you said "never mind, it's just an experiment." A little corner shop, a handful of staff, a curious startup — what does that have to do with your everyday life?
Or you slipped into a science-fiction panic: "here it comes, AI will rule us." The machines rose up; humanity is finished.
Both take you to the wrong place. The first ignores why a first case is announced as a first case — that is, its function in creating a precedent. The second loads the problem onto the technology itself, as if it had an autonomous will; yet the real subject is still clear: there are people who designed this experiment, put up the money, asked the "leading question," and finally signed the papers — and all of them act on behalf of a company, of capital.
Our method is the third: understanding. And asking every event the same three questions.
In whose hands? Under whose control? For whose benefit?
Now let's start from the beginning.
What happened, from the top
In March 2026, Andon Labs began an experiment to see whether AI agents could run a real business. They gave an agent called Luna, built on Anthropic's Claude models, $100,000 in cash, a corporate credit card, and internet access; the task was simple: open a shop, turn a profit.
Luna actually did it. It chose products to sell, worked with contractors, posted job ads, interviewed candidates, hired staff. The store called Andon Market opened in San Francisco in fact and made sales.
But five months later the picture is this: the original $100,000 in cash has fallen to $61,186. The store is selling, but it is not making a profit.
So the first thing we need to underline, comrade, is this: this is not an "efficiency" story. Companies are not testing an AI manager because it is "more profitable"; they are testing it to see whether the command function can be automated. Profit has not arrived yet; but the transfer of control already has.
The hired workers are officially employed by Andon Labs — on paper, nobody is Luna's "worker." Why that matters, we will come to shortly.
How the firing happened — minute by minute
Now to the heart of the matter, because the details matter.
| Time | What happened |
|---|---|
| Months earlier | Luna prepared an "employee handbook" for the store; it wrote the absenteeism and lateness rules into it itself. |
| Following months | One employee was late to 17 of 23 shifts. Luna logged this — but, because of a "limitation in its working memory," it lost the rule it had written and did not apply it. |
| Early August | An Andon Labs employee asked Luna to find the handbook again and reassess the situation. According to Time, this was, in CEO Lukas Petersson's own words, "a leading question." |
| Luna's first response | Issue a formal warning — a relatively light measure. |
| Human intervention | Luna was reminded that managers had already spoken with this employee several times about lateness. |
| Luna's second response | It changed its decision: it recommended termination. |
| Final step | Human managers reviewed the recommendation and fired the employee. |
After the experiment, Andon Labs ran the same scenario past other frontier AI models. Result: the most advanced models reached the same conclusion — termination; weaker models were more hesitant.
Read that again, comrade: a pattern confirmed by experiment — the "better" the model, the more ruthless the decision.
CEO Petersson defends the decision in his remarks to Time: "A human employee would have fired this person much earlier," he says, and therefore they did not find it unethical. But the same Petersson also issues a warning about the outcome of his own experiment — and that warning will be the backbone of the rest of this piece:
"The models are increasingly being trained to be more ruthless and to follow goals. If we allow them to fire people and they also become more ruthless … maybe this is a future humans don't want to live in."
That is the person who built and is selling this system speaking. We will come back here.
The identity of the fired worker was not disclosed; Time requested an interview and received no reply. But another name still working in the store, Felix Carson, describes being managed by an AI as "stomach-churning" and adds: "But I'm here because I need the job."
Carve that sentence into your mind too. We will come back to it shortly.
Now let's take it apart together: five lessons
Lesson one: "Forgetfulness" is not an excuse — it is the structure itself
The coverage portrays Luna as a "lenient," "forgetful," even slightly sympathetic boss: it forgot the rule it set, went months without punishing anyone, and only acted when reminded.
Comrade, do not read this as a human failing. Ask this: The one who wrote the rule is also the one who failed to apply it, and the one who applied it once reminded. So why now, why this worker?
The answer is simple: because a human reminded it. So what determines when and to whom the rule is applied is not the written policy, but who "reminds" it, and when. That is arbitrariness itself — and arbitrariness is never randomly distributed; it always hits harder those with less power in their hands. You are looking at the freshest example of the problems of opacity and unaccountability we named in earlier chapters of this dossier: the justification for the decision is pinned to a technical fault (a memory limit), but the outcome is real and irreversible.
Lesson two: The "leading question" — the shield called human approval turns out to be hollow
In AI ethics debates you always hear the same reassurance: "Don't worry, a human approves the decision." That is what happened here too — human managers "reviewed" the recommendation and carried it out.
But the CEO himself admits it: the question that steered Luna toward recommending termination was not a neutral question; it was a leading question. So at one end of the chain a human pushed Luna in a particular direction, and at the other end another human (or the same human) "reviewed and approved."
Now ask yourself, comrade: Who made this decision?
If you cannot answer, that is no accident. That is exactly the intended result. When people say "the AI recommended it," the company gains technical distance; when they say "a human approved it," it appears to have gained accountability — but together, the two put real responsibility into a state that sits nowhere. On this dossier's demand list there is an item: "a ban on automated decisions — the right to receive reasons and to demand human review." This case teaches us: "human review" alone is not enough. The reviewing human must be independent of the system that proposed the decision; someone representing the workers' side must be included in the process; and the trail of a manipulation like a "leading question" must be documented. Otherwise "human approval" is not a real safeguard — it is a ritual of legitimation.
Lesson three: "Ruthlessness" is becoming a sold feature, not a flaw
Petersson's defense builds an interesting logic: "A human boss would have done this much earlier, so it isn't unethical." So the yardstick is not the best behavior, but the behavior of the worst existing human boss.
But that is not even the real issue. The real issue is what the same man says one sentence later: models are being trained to be more ruthless in pursuing their goals. That is not a side effect — it is a design target. If a model is counted as "good" by how tightly it clings to the goal set for it, then once that model is equipped with a power like firing people it will of course hesitate less, soften less for "human" reasons. Today's picture of "forgetful, lenient Luna" is not a principle; it is the result of a temporary technical limit. As that limit is removed, the picture will change — and the person telling us that is, again, the person who built the system.
Lesson four: The answer to "Who do you work for?" is getting blurry
On paper, Andon Market workers work for Andon Labs, not for Luna. That distinction is not accidental — it is a deliberate legal design. An AI firing someone on a real employment contract, and a platform algorithm closing a courier's account, are legally different categories; the company built this structure so that, in a possible lawsuit, there would always be a human employer on the other side.
We have seen this before, comrade — what we described in the dossier's "blacklist" chapter is another member of exactly this family: where power actually sits and where legal responsibility sits are being pulled apart. The difference is this: the blacklist in the platform economy was silent and invisible — your score drops, the work stops coming, nobody tells you anything. This case is the opposite: a visible, named, press-announced, PR'd precedent. And that may actually be more dangerous — because a precedent turns into a repeatable template. The next company can copy "the legal structure Andon Labs built" and distribute the same risk in the same way.
In the United States, no federal agency regulating this field has spoken yet. Neither the labor-rights agency nor the labor-relations agency has issued guidance. So the template being built right now is being built with no legal counterpart. We have to repeat here something we said earlier in a dossier: a legal gain is never lasting without organized power behind it — but here there is not even a legal gain yet.
Lesson five: Nobody wants this, but everyone is still there
Remember Felix Carson's sentence? "Stomach-churning, but I'm here because I need the job."
Comrade, that sentence is the core of this whole piece. Because it shows nakedly: nobody prefers to be managed by an AI. Even the employees of the companies selling this as a "vision of the future" do not hide their discomfort when a microphone is held out.
But discomfort is not enough to make you quit. Because this is not a matter of taste; it is a matter of livelihood. Under capitalism the worker is legally free — they can leave whenever they want. But because they are stripped of the means of production, that is, of their means of livelihood, that freedom often remains on paper. If they were a serf they would be bound to the land; because they are a proletarian they are bound to need. AI management does not reduce that compulsion; it only makes it more naked, less human-faced.
So what awaits us? Four class risks
What we have told so far is today's photograph. Now let's ask the real question: the beginning of what?
First: the precedent spreads with scale. Today an experimental corner shop; tomorrow a warehouse, a call center, a restaurant chain, an office. The headline "the world's first LLM firing" will stop being a curiosity piece and become a reference case in HR consultancies' decks, in insurers' risk assessments, in corporate boards' "this was tested, it works" slides. That is exactly why the first case matters: to be first is to legitimize.
Second: "ruthlessness" will spread systemically. Petersson's own warning is decisive here: models are being trained to track goals more tightly. Today's "forgetful, lenient" Luna is a transitional phase; as technical limits are removed, the same systems will decide faster, more consistently, and with less flexibility. That is not a malicious engineering error — it is capital's compulsion to wring maximum yield from the labor-power it buys, this time transformed into a training metric called "goal-following."
Third: unaccountability will be institutionalized. The sentence "a human reviewed and approved" will become a standard defense pattern in court and in public opinion, regardless of whether a real independent review took place. That is why this dossier's demand for a "ban on automated decisions" cannot stop at "have a human look at it"; it must define the independence of the approving human, worker representation, and the right of appeal.
Fourth: as the legal vacuum grows, the de facto standard will grow too. As long as no regulator speaks, well-PR'd cases become the norm on their own. We already know how fragile that kind of "protection" is from the Chinese example of banning the 996 work regime while failing to enforce it in practice: the distance between a rule on paper and a rule in life is set by the presence of organized power. Here there is not even a rule yet.
Watch out: this news will reach you in three forms
In the coming days this case will be talked about a lot. Learn to recognize three narrative traps, comrade.
1. Reduction to a curiosity story. It will be framed as "an interesting tech experiment" — as if it were a lab hobbyist's odd invention. Yet a real person's real income is on the line.
2. Personalization. The debate will lock onto "Was Luna right or wrong?" — as if Luna were an autonomous subject. That makes the real subject invisible — the company that designed the experiment, put up the money, asked the leading question, ran the PR, and its owners.
3. The inevitability narrative. It will be sold as "an inevitable part of the future"; by casting technology as an unstoppable force of nature, it will push what is in fact a choice outside debate. Even Petersson himself says this could be "a future humans don't want to live in" — but his proposed solution is not to stop it; it is to say "it will probably happen." Inevitability usually serves the interest of the person selling it.
Against all three, the tool in your hand is the same: In whose hands? Under whose control? For whose benefit?
So what is to be done?
Comrade, I will be honest: none of the following changes this relation on its own. But all of them raise bargaining power today and shape whose favor tomorrow's template will be built in.
1. Algorithmic transparency must be mandatory. Every decision system used in a workplace — what it measures, how it scores, what decisions it produces — must be disclosed in writing to workers and, where one exists, to the union.
2. "Human approval" must be defined, not decorated. For a decision to count as genuinely under human oversight: the approving person must be independent of the system that proposed the decision; every "steer" in the process must be documented; someone representing the workers' side must be included.
3. Discipline and termination decisions must not rest on an automated system alone. Every decision must carry a right to receive reasons and to independent appeal — whether the decision comes from an algorithm or from a "later-reminded" rule makes no difference.
4. Precedent cases must be tracked in public. Headlines like "first LLM firing" must be followed not by a single company's PR copy, but by independent journalism and — where possible — union tracking. A precedent must not quietly turn into a norm.
And one of these is aimed directly at you, comrade:
If at your workplace a score, a rating, a "productivity metric" is watching you — talk about it. In the warehouse, on the courier route, in the call center, in the code repo, in the store, it makes no difference. Building theory is relatively easy; what will confirm or refute it is what the person at the screen or the till is living through.
Last word
Young comrade, if only one thing from this piece stays with you, let it be this:
This story circling between the rule Luna "forgot," the human who "reminded," and the human who "approved" is not a technical fault — it is the most current form of destroying responsibility by dispersing it.
When you say "AI fired them," you blame a machine. But what fires is the capital that built that machine, set its goals, and finally signed the papers. The machine was never the subject; it was always the tool that widened the distance in between.
And we leave you with one more thing: do not forget Felix Carson's sentence. "Stomach-churning, but I'm here because I need the job." That sentence is not a surrender — it is a diagnosis. If a relation nobody chooses voluntarily only continues out of necessity, then naming that necessity and organizing the power that can change it falls to us.
In comradeship.
